Why Marketing Mix Models (MMM) are Ambitious
Effective media planning needs to accommodate the complex variables that impact your target audience and drive sales.
MARK GARRATT – Marketing mix modeling is a very ambitious type of analysis – it seeks to provide a single, holistic view of all the drivers of sales. The purchase behavior that we are trying to model may have many causes. Some of those causes are due to advertising and promotion and some are due to uncontrollable factors like the weather or the economy. Within the advertising-driven behaviors, there is another major dimension of difference: where does the advertising act in the consumer journey? Does advertising act at the top of the funnel generating broad awareness and reach, or does it act at the bottom of the funnel, capturing demand through timely and clever interventions? Linear TV is a typical top-of-funnel medium; retail media (such as the search box on Target.com) is one of the newer lower-funnel or demand-capture media.
Yet another dimension that a marketing mix model (MMM) must address is whether an ad is acting directly or indirectly. For instance, a display ad that is targeted to your phone is direct. But a TV spot that advertises a web domain is acting indirectly. MMM is ambitious because it seeks to capture all these impacts, from the most general to the most particular, in one model framework.
If we think of the major ways that data inputs can vary in a Marketing Mix Model (MMM), then we can identify several factors:
- Controllable versus non-controllable
- Upper, middle, or lower funnel
- Direct or indirect
- Whether costs can be defined or not
- The geography in which the advertising is planned or acts
These factors can apply equally to digital or linear media driving sales in brick & mortar or online classes of trade. The typical models used for MMM are based on regression. Regression is flexible. Nothing prevents us from trying to sum up the data to a fixed geography (store, chain, household, individual) in a fixed time (day, week, month) and put all these different types of variables in one model. However, the model does not “know” things that you might take for granted. It does not know how the factors described above may distort the impact of advertising. Here are two examples of how complexity can find its way into models:
- When it’s raining, I don’t want to go outdoors, especially if it’s cold or stormy. If I am trying to get dinner for the family, I may be more inclined to use a delivery service like GrubHub or UberEats. If I use one of those services, I will see ads for restaurants that may change where I end up placing my order.
- As an advertiser, I spend part of my budget on social media. Part of that social media is national, but another part is market-level. Over time, I have shifted money from national social to market-level social. I have also used that market-level social specifically to introduce a loyalty app.
In the first case, we are dealing with an uncontrollable but important factor (the weather) that makes me shift from one distribution channel to another (a change in geography/place). While in that distribution channel, the consumer gets exposed to retail media (which may be originated by the retailer, not by the advertiser) which results in a purchase different from their regular brand.
In the second case, the advertiser is changing the weight of the impression on social media but that really is a shift to market-level (geography) which also comes at a higher cost. That depresses their short-term ROI. The local social media is operating at mid-funnel level, but it is being used to launch an app that is at lower-funnel or demand-capture level. The app is designed to bring the customer back to the store to get more “share of wallet” (an indirect effect) and part of the cost of market-level social media needs to be traded off against building customer loyalty with the app.
The purpose of these examples is to show just how common it is for the impact of intended marketing effects to get amplified, diluted, or changed, sometimes by things out of our control (like the weather) and sometimes by the unintended consequences of our own plans. Also, it may begin to show the way that just putting more and more variables into regression models and expecting to get true reads on contribution to sales might be naïve. There are specific interactions. There are effects at different levels of action. Costs keep changing. And there are a lot of potential effects.
How Hierarchical Bayes Helps Us Handle Ambitious Models
Bayesian analytics was named after the Rev. Thomas Bayes 1701-1761, who Brittanica describes as an “English nonconformist theologian and mathematician who was the first to use probability inductively and who established a mathematical basis for probability inference (a means of calculating, from the frequency with which an event has occurred in prior trials, the probability that it will occur in future trials.”
Such theories couldn’t be practically applied until the 1990s when major advances in computation made practical versions of it possible. Hierarchical Bayes describes a model that has layers or hierarchies; most often those layers are defined by different geographies or by-product characteristics. As described above, Bayesian models also have “priors”. Priors (which simply put are preliminary assumptions that may come from experience or data) are a useful way to build domain knowledge such as sign constraints into the results.
Hierarchical Bayes Regression allows us to construct models where we can handle many of the problems that make MMM tricky:
- They allow us to break up the data in hierarchies of granularity. So that store-level effects (like local weather) can act at the store-level, DMA-level effects (like terrestrial radio) can act at the DMA level and national media (like national social media) can act at a national level. This ability to break up the data into levels of action solves many problems of misattribution.
- They allow us to apply priors to variables so that domain knowledge derived from external norms, known constraints (e.g., advertising effect is positive), MTA findings, A/B tests, etc. is incorporated into results. This ability gives answers that make sense, are consistent over time, and lead to KPIs that drive action through sequential measurement.
- But even more important, they are self-norming, so that if you have multiple products, packages, outlets, etc. pooled together, they can share the measure of advertising impact across similar products or geographies and reinforce the granular estimates. And if you have outliers, their values are guided toward the central tendency, preventing mistakes and phony anecdotes.
- They are robust when media (or any effect) applies to only a subset of products or distribution channels (e.g., this retail promotion only applies to Target sales). This addresses the complexity in media that we see especially in the evolving online space.
- Because they operate flexibly at granularity, Bayesian models allow us to measure many more variables than simpler forms of regression executed piecemeal.
- They are efficient. Because when they are done right, Bayesian models offer high levels of control against “random acts of data.” So, analyst time can shift from babysitting models to interpretation of results and implications for action.
As we said at the start, MMM analytics are ambitious and complex. To get them right you need the kind of analytics partner who is an expert at wielding the sharpest tools (we would argue Hierarchical Bayes) to get you actionable recommendations for efficiency in media spend. Your analytics partner also needs to be knowledgeable of the evolving media landscape to demystify and decipher the real contributions of any media channel to your growth objectives.
In future posts, we will give you our POV on how to account for retail media when looking at your overall mix as well as topics like the challenges and opportunities modeling the impact of PR, affiliate, and influencer marketing against all your other marketing investments.
If you have a critical topic that is impacting your media mix decisions and would like to know more about how we can apply our special expertise to the advantage of your brand, let’s have a conversation. We would love to help – contact us at email@example.com.
Mark Garratt is a partner and co-founder of in4mation insights. He is an accomplished analytics professional with a distinguished career in both business and academia. Mark has been a trusted advisor to some of the world’s biggest brands and an analytics leader at CPG companies including P&G, SABMiller Brewing, and The Gillette Company.